A Comparison of Variable Selection Methods in Competing Risks Models for Business Failures

The statistical techniques that are applied to the analysis and prediction of the failure of business enterprises include regression modelling of hazard functions. Because a large number of financial variables are available as potential predictors, it is necessary to choose an appropriate variable selection procedure in this modelling. Furthermore, alternative models, for cause-specific and subdistribution hazards, are available when different modes of failure are being examined as competing risks. We demonstrate the application of both types of hazards to the study of failures from bankruptcy, liquidation or dissolution among 75479 manufacturing firms in seven European countries in 2000-2018, and we examine the results of variable selection by stepwise and lasso methods for both approaches. All analyses resulted in successful prediction, with areas under the ROC curve of 0.90 or just below, but the lasso approach achieved this with a smaller number of variables selected.